Research On Person Counting Method Based On Improved YOLOv3

With the advancement of urbanization,the flow of people has become more frequent.The analysis and statistics of crowd activities is an urgent need in real scenes such as scenic spots,campuses,workplaces,and transportation hubs.Recognizing people from images and making statistics is an economic and convenient means to achieve this demand.Target detection and recognition is an important subject in the field of computer vision technology.Existing algorithms and solutions have problems such as inaccurate statistics and high thresholds for people counting.In response to these problems,this thesis proposes a convolutional neural network model PCS-YOLOv3(PCS for short)for people counting.PCS has improved the target detection model of YOLOv3.According to the characteristics of people counting tasks,it adds to the original neural network structure Three shortcuts are used to enhance the ability of the unbalanced branch with a deeper number of layers in the original model to obtain the characteristics of the residual network,thus constructing a convolutional neural network model specially constructed for people counting.In order to make this model have a more accurate effect on people statistics,this thesis proposes a method of using 3D engine rendering to generate training samples,and use this as a data set to train the neural network.Experimental tests show that the PCS model is more accurate than YOLOv3 and other target detection models on the task of people counting.The modification of the neural network structure and the 3D generation of training samples both have gains in the results.This thesis uses the Flask network framework technology to implement the API and front-end web page for people counting by calling the trained neural network model,making this method easier to use.